Int J Biometeorol (2012) 56:357–370
DOI 10.1007/s00484-011-0439-0
ORIGINAL PAPER
Measuring the impact of temperature changes on the wine
production in the Douro Region using the short time fourier
transform
Mário Cunha & Christian Richter
Received: 11 July 2010 / Revised: 9 January 2011 / Accepted: 12 February 2011 / Published online: 15 May 2011
# ISB 2011
Abstract This paper investigates the cyclical behaviour of
the wine production in Douro region during the period
1932–2008. In general, wine production is characterised by
large fluctuations which are composed of short-term and/or
long-term cycles. The aim of this paper is twofold: firstly,
we decompose the wine production's variance in order to
find the dominating production cycles, i.e we try to explain
whether wine production follows more long-term or short-
term cycles. In the next step, we try to explain those cycles
using a dependent variable, namely the medium spring
temperature (Tm_Sp) for the period 1967–2008. We
estimated a Time-Varying Autoregressive Model, which
could explain 75% of the production that is characterised
by 4.8- and 2.5-year cycles. We use the Short Time Fourier
Transform to decompose the link between wine production
and temperature. When the temperature was incorporated,
the R2 increased and the Akaike criterion value was lower.
Hence, Tm_Sp causes a large amount of these cycles and
the wine production variation reflects this relationship. In
addition to an upward trend, there is a clearly identifiable
cycle around the long-term trend in production. We also
show how much of the production cycle and what cycle in
particular is explained by the Tm_Sp. There is a stable but
M. Cunha (*)
Department of Geociências, Ambiente e Ordenamento do
território, Faculdade de Ciências, Universidade do Porto,
Campus Agrário de Vairão, Rua Padre Armando Quintas,
4485–661 Vairão, Portugal
e-mail: mcunha@mail.icav.up.pt
URL: http://webpages.icav.up.pt/pessoas/mccunha/
C. Richter
Royal Docks Business School of Economics,
University of East London,
London E16 2RD, UK
e-mail: c.richter@uel.ac.uk
not constant link between production and the Tm_Sp. In
particular, the temperature is responsible for 5.2- and 2.4-
year cycles which has been happening since the 1980s. The
Tm_Sp can also be used as an indicator for the 4.8- and
2.5-year cycles of production. The developed model
suggests that stationarity is a questionable assumption,
and this means that historical distributions of wine
production are going to need dynamic updating.
Keywords Spectral analysis . Time-varying spectra .
Kalman filter . Wine production modelling . Climate
variability
Introduction
Regional wine production is characterised by large inter-
annual fluctuations with adverse consequences for every-
body involved in vineyards and the wine industry as a
whole. Moreover, the impact of these fluctuations have
consequences concerning input use, land use and thus,
indirectly, the environment.
Moreover, Santos et al. (2010) and Esteves and Orgaz
(2001) investigated the impact of the climate on wine
production in Portugal, that is in the Douro and Dao
regions. Both studies found a significant impact of the
climate on the wine production. In the Mediterranean
region, climate is a major risk, probably due to the broad
spectrum of possible negative climate events and to the
uncontrolled aspects of natural hazards (Quiroga and
Iglesias 2009). The main part of the inter-annual variability
in the atmospheric circulation of the western European
(Iberian) areas can be tied to large-scale geophysical
mechanisms of which the most prominent is the North
Atlantic Oscillation (NAO) (Hurrell 1995; Hurrell and Van
358
Loon 1997; Corte-Real et al. 1998). Through its control
over regional temperature and precipitation variability,
atmospheric teleconnections can affect vegetation dynamics
(Vicente-Serrano and Heredia-Laclaustra 2004) and wine
productivity and quality (Esteves and Orgaz 2001). In this
context, climate regulates nearly every step of wine
production from selection of a suitable grape variety to
the type and quality of wines produced.
Detecting and characterising changes of the dynamics
involved in wine production over time is the natural first
step towards identifying the key determinants of this
process. Various crop modelling approaches are proposed
in the agricultural literature focusing on investigating the
cyclical factors associated with crop production and
studying how these crop cycles should be characterised
and measured. In order to support management decisions, a
number of parametric and non-parametric risk analysis
models were developed (Ferris 2006; Goodwin and Ker
1998; Capitanio and Adinolfi 2009; Gemmil 1978; Chen
and Chang 2005). However, there are only a few risk
analysis models available for wine production (Folwell et
al. 1994; Quiroga and Iglesias 2009). Usually, such
analyses are based on probability distributions arising from
historical data where the distributions developed are based
on at least a partial assumption on stationarity and do not
incorporate sensitivity tests or (estimate) changes in
distribution outcomes (Chen et al. 2004; McCarl et al.
2008; Zhu et al. 2008). Evidence exists that climate
change will shift the mean and variance of crop yields,
challenging the stationary assumption (Milly et al. 2008).
In this context, crop yield-related risk analysis would need
to use distributions with non-stationary means and varian-
ces along with possibly shifting higher order moments
(McCarl et al. 2008).
The link between wine production and in-season climate
data has been studied through many research articles (Folwell
et al. 1994; Cunha et al. 2003; Bindi et al. 1996). However,
the perennial nature of grapevine means that every year is
physiologically dependent in several ways on the previous
years (May 2004; Vasconcelos et al. 2009). The permanent
structures of grapevine provide reserves of carbon and
nutrients but also carry effects from year to year. Therefore,
developing multiple years’ models is very difficult as many
of the important carry-over effects on growth or cropping are
not well understood.
The traditional approach to modelling the cyclicality of
wine production is to assume normal climate and project
crop yields as a linear extension of past trends. Although
this linear trend describes the overall long-term trend in
production, it does not reflect the inherent cyclicality in
production and the information contained therein. Indeed,
the explanatory power of a linear model is less than 50%
(Phares 2000; European Commission 1997; Santos et al.
Int J Biometeorol (2012) 56:357–370
2010). This implies that, instead of using this model to
forecast wine production, one could as well flip a coin.
When conducting a long-time cyclical analysis, one has to
take into account that the analysis is affected by structural and
climatic variability of the vineyard. At this time there is no
regional analysis that co-integrates these two sources of
information.These are the changes we wish to test for here.
Enhanced production and climate effects will come in several
parts: obviously, climate affects production via grapevine
growth convergence (coherence, correlation), but production
methods have changed themselves and there is a non-constant
impact (or spillovers) of climate onto wine production, which
translates into stronger lead/lag relationships between pro-
duction and climate. We examine all three in this context,
focusing on measures of coherence and gain, respectively. We
can then ask: to what extent are production cycles becoming
more correlated to climate?
We are therefore engaged in an exercise in identifying the
linkages between climate and wine production using a time–
frequency approach. We are not aware that this has been
attempted before. In this paper, we show how to use a time-
varying spectral analysis to determine the degree of impact at
different frequencies and cycles, even where data samples are
small and where structural breaks and changing structures are
a part of the story. The inconclusive results obtained in the
past may have been the result of using a correlation analysis
which averages the degree of contemporaneous impact across
all frequencies. That is problematic because two variables
could share a trend or short-term shocks, but show no
coherence between their cycles. That would imply low or
possibly negative contemporaneous correlations, and give no
picture of the true linkage or dependence between them.
A common feature of all the studies cited above is that the
results are sensitive to: (1) the choice of coherence measure
(correlation, concordance index); (2) the choice of cyclical
measure (classical, deviation or growth cycles); and (3) the
detrending measure used (linear, Hodrick-Prescott filter,
band pass, etc.). This sensitivity to the detrending technique
is a serious difficulty highlighted in particular by Canova
and Dellas (1993) and Canova (1998). The advantages of
using a time–frequency approach are therefore:
– It does not depend on any particular detrending
technique, so we are free of the lack of robustness
found in many recent studies. These methods also do
not have an “end-point problem”
– No future information is used, implied or required as in
band-pass or trend projection methods.
– There is no arbitrary smoothing parameter, such as in
the HP algorithm, equivalent to an arbitrary band-pass
selection (Artis et al. 2004).
– We use a coherence measure which generalises the
conventional correlation and concordance measures.
Int J Biometeorol (2012) 56:357–370
Any spectral approach is tied to a model based on a
weighted sum of sine and cosine functions. However, that is
not restrictive. Any periodic function may be approximated
arbitrarily over its entire range, and not just around a
particular point, by its Fourier expansion (a suitably
weighted sum of sine and cosine terms), and that includes
non-differentiable functions, discontinuities and step func-
tions. Hence, once we have time-varying weights, we can
get almost any cyclical shape we want. Therefore, a time-
varying spectral approach, capable of separating out
changes at different cyclical frequencies in the regional
wine production, will be needed to provide the flexibility to
capture these features. Similarly, a time-varying approach
will be necessary if we are to accommodate the structural
breaks which must be expected with wine production
methods and meteorological variability. However, if these
changes argue for a time-varying approach to measuring the
coherence between variables, then they also argue for a
decomposition of the different cycles that make up wine
production performance. Hence, our choice of a time–
frequency approach.
Materials and methods
Wine industry, Douro region and sample data
Portugal is number five in the European wine producers
ranking and the number ten world-wide (OIV 2010), and it
accounted for 7% (246 103 ha) vineyard surface and around
4% (7.1 106 hL) of the total wine production in the EU-15.
The Douro Region, located in northeast Portugal, has an
area of 250,000 ha and vineyards cover approximately 15.4%
of all the land in the region. Viticulture, the main activity of
most farmers in the region, takes place under particularly
rigorous climatic conditions, on stony soil that cannot be put
to any other use. The Douro region is characterised by having
mountain viticulture, with more than 70% of vineyards
planted on hillsides with a slope greater than 30% (IVDP
2010). The most noteworthy red wine varietals, all of them
native to the region, are Touriga Franca, Touriga Nacional
and Tinta Roriz (tempranillo).
Out of the entire amount of land used as vineyards in the
Douro region, only 26,000 ha (about 68%) are authorised
for the production of Port Wine. The vines which are
considered appropriate for this wine type are selected
according to a criteria of quality based on a scoring method
(considers soil, climatic, varieties, age of the vines), and
classified according to a scale of quality that ranges from A
to F.The importance of climate conditions for wine
production in Douro is emphasised by the score leading to
the A–F classification were the parameter related with
climate conditions represents 62.5% (Fonseca 1949).
359
The regulating institution of the sector (IVDP) stipulates
annually the quantity of grapes produced in Douro Region
that can be used for the production of Port wine. This
typical quota policy, called “benefício” coefficient, is
attributed annually among the vineyard that have that
property right (classification A to F).
The meteorological observations for the years 1967–
2008 were collected in station of Peso da Régua (41º10′N,
7º47′W, 139 m above sea level), located within the Douro
region and has not been relocated over the period of record.
The meteorological data consist of daily observations mean
temperature (Tm; °C) and precipitation (R; mm), that are
summed (R) or averaged (Tm) by seasons (spring and
summer).
According to Köppen’s classification, the region belongs
to group Csb (temperate, with dry summer, which is not
very hot but extensive), while Thornthwaite’s rational
climate classification describes it as B1B’2s2a’ humid
(hydric index: 25.3%; B1), mesothermic (thermal efficiency
index: 778 mm evapotranspiration; B’2), with great shortage
of water in the summer (aridity index: 38%; s2) and thermal
efficiency summer concentration index: 47% (above 20% =
typically continental).
The Douro‘s vineyards is one of the most non-irrigated
arid regions of the Europe and a strong water stress is
normally observed. These situations are especially frequent
in summer and appear as a consequence of the low soil
water content (stony soils), due to the low rainfall and the
elevated gradients of the water vapour pressure between the
leaves and the air (Chaves and Rodrigues 1987). It is well
established that positive NOA is associated to a decrease in
moisture conditions and drought episodes in southern
Europe and in Mediterranean areas (Corte-Real et al.
1998; Knippertz et al. 2003) with great impact for the
Douro Valley (Paredes et al. 2006).
The mean annual precipitation in Douro region vary from
400 to 900 mm and the mean monthly temperatures range
from 5 to 8°C (January) up to 21–24°C (July). During the
period April–October, the mean temperature is about 19.5°C
and according to the climate maturity grouping (Jones 2007),
the growing season can be defined as “warm” (April–
October; mean temperature between 19 and 24°C). In the
ripening period (20 July to 20 September), the rainfall in
80% of years is ≤28 mm (Reis and Lamelas 1988) and the
available water reserve at the end of the ripening period is
always ≤20%, causing lower berry weight and, consequently,
lower wine yield (Cunha et al 2003).
In this work, we use the annual wine production data
(1932–2008) for the Douro Region provided by the
Instituto dos Vinhos do Douro e Porto (IVDP 2010).
The regional wine yield in any given year is the net
effect of all the planting decisions (technologies include)
that modify both the regional vineyard area and the age
360
composition of grapevine composition stock. In Douro
regions, the drawback of wine yield estimations is the
absolute lack of annual detailed regional data of the
dynamics of new plantings, replanting, removal and age
composition of vineyard. However, the percentage share of
young vineyards (less than 4 years) in 1998 is less than 4%
in comparison to all other vineyards (Casa-do-Douro 1999).
In this context, our modelling assumption is that this share
remained constant in our sample and is reasonably small
not to jeopardise the stability of the productivity link.
Empirical techniques
Estimation in the time domain
We estimate the bilateral links between the wine production
cycles. In order to allow for the possible changes in the
parameters, we will employ a time-varying model AR(p) by
applying a Kalman filter to the chosen model as follows:
X
9
yt ¼ a0;t þ ai;t yti þ "t ð2:1Þ
i¼1
with
ai;t ¼ ai;t1 þ hi;t ; for i ¼ 0:::9 ð2:2Þ
 
and "t ; hi;t  i:i:d: 0; s 2";hi ; for i ¼ 0:::9.
In order to run the Kalman filter, we need initial
parameter values. The initial parameter values are obtained
estimating them by Ordinary Least Squares (OLS) using the
entire sample (see also Wells 1996). Of course, using the
entire sample implies that we neglect possible structural
breaks. The initial estimates might therefore be biased. The
Kalman filter, however, corrects for this bias since, as Wells
(1996) shows, the Kalman filter will converge to the true
values independently of the initial values. Hence, our start
values have no effect on the parameter estimates, i.e our
results are robust. Given these starting values, we can then
estimate the parameter values using the Kalman filter. We
then employ a general to specific approach to obtain a final
specification for (Eq. 2.1), eliminating insignificant lags
using the strategy specified in the next paragraph below.
The maximum number of lags was determined by the
Akaike Criterion (AIC). The AIC takes indirectly into
account whether a variable is significant or not. If not, then
the AIC value usually drops. Each time we ran a new
regression, we used a new set of initial parameter values.
Then, for each regression, we applied a set of diagnostic
tests, shown in the tables in the following sections, to
confirm the final specification found. The final parameter
values are therefore filtered estimates, independent of their
starting values.
Int J Biometeorol (2012) 56:357–370
Using the specification above implies that we get a set of
parameter values for each point in time. Hence, a particular
parameter could be significant for all points in time, or at
some periods but not others, or it might never be
significant. These parameter changes are at the heart of
this paper as they imply changes in the lag structure and
hence changes in the spectral results. If a parameter was
significant for some periods but not others, it was kept in
the equation with a parameter value of zero for those
periods in which it was nonsignificant. This strategy
minimised the AIC criterion, and led to a parsimonious
specification. Finally, we tested the residuals in each
regression for auto-correlation and heteroscedasticity.
The final specification (Eq. 2.1 and 2.2) was then
validated using two different stability tests. Both tests
check for the same null hypothesis [in our case a stable AR
(9) specification] against differing temporal instabilities.
The first is the fluctuations test of Ploberger et al. (1989),
which detects discrete breaks at any point in time in the
coefficients of a (possibly dynamic) regression. The second
test is due to LaMotte and McWorther (1978), and is
designed specifically to detect random parameter variation
of a specific unit root form (our specification). We found
that the random walk hypothesis for the parameters was
justified for each model (results available on request). We
also test for autocorrelation of the residuals. For this
purpose, we use the Ljung-Box test, which allows for
autocorrelated residuals of order p. In all our regressions,
we could reject the hypothesis of autocorrelation.
Finally, we chose the fluctuations test for detecting
structural breaks because the Kalman filter allows for
structural breaks at any point and the fluctuations test is
able to accommodate this. It should be noted that all our
tests of significance, and significant differences in param-
eters, are being conducted in the time domain, before
transferring to the frequency domain. This is because no
statistical tests exist for calculated spectra (the data trans-
formations are nonlinear and involve complex arithmetic).
Stability tests are important here because our spectra are
sensitive to changes in the underlying parameters. But,
given the extensive stability and specification tests con-
ducted, we know there is no reason to switch to another
model that fails to pass those tests.
Once this regression is done, it gives us a time-varying
AR(p) model. From this AR(p), we can then calculate the
short time Fourier transform as outlined below, and as
originally suggested by Gabor (1946), in order to calculate
the associated time-varying spectrum.
Spectrum analysis
As a first step, we analyse the power spectral density
function of the wine production in Douro region. The
Int J Biometeorol (2012) 56:357–370
power spectral density function (PSD) shows the strength
of the variations (energy) of a time series at each frequency
of oscillation. In other words, it decomposes the variance of
a time series into its periodicities. In a diagram, it shows at
which frequency variations are strong/powerful, and at
which frequencies the variations are weak (expressed in
“energy”). The unit of measurement in the PSD is energy
(variance) per frequency, frequency band or cycle length.
For example, if a time series Xt ¼ "t ; where "t 
i:i:d:ð0; s 2 Þ and constant over time, the power spectrum
would look like Fig. 1.
As one can see from Fig. 1, a white noise process is
characterised by the fact that no specific frequency has a
bigger impact than any other frequency, for w ¼ 0; :::; p.
However, if the data were dominated by long production
cycles, then the diagram would have higher power
(variances) at the low or middle frequency bands respec-
tively, and lower power at the high frequencies.1
In order to calculate the spectrum from an estimated
representation of Eq. 2.1, we use the Fast Fourier
Transform. The Fast Fourier Transform is an efficient
algorithm for computing a discrete Fourier transformation
or in our case a Discrete Time Fourier Transform (DTFT)
for discrete points in time. In our case, it creates a frequency
domain representation of the original time domain repre-
sentation of the data (Eq. 2.1). Hence, our analysis of the
spectra and coherences that follow are based on a
regression done in the time domain, but then transformed
into a frequency domain function by the Fourier transform.
However, in this paper, we also allow the coefficients of our
regressions to vary over time. Therefore, we derive one
DTFT for each point in time. For technical details, please
refer to the Appendix.
Thus, when we present our empirical results below, they
are based on the time-varying STFT calculations. The only
Energy
2 employ the fluctuations test (Ploberger et al. 1989). Finally,
Frequency ( )
Fig. 1 Power spectrum of a white noise process
1
In the spectral diagrams that follow, we use the term “power” rather
than “energy” to denote relative variances.
361
difference from Fig. 1 is that we have to add a time
dimension to show how the spectra have changed over
time. The result is then a 3-dimensional diagram.
Cross-spectrum analysis
In this paper, we also investigate the linkage between
different wine production cycles. In the frequency domain,
the natural tool to do that is the coherence. The spectral
coherence KXY 2
is a statistic that can be used to examine
the relation between two signals or datasets. Values of the
coherence will always satisfy0  KXY 2
 1. For a strictly
proportional linear system with a single input xt and single
output yt, the coherence will equal one. If xt and yt are
completely unrelated then the coherence will be zero. If
2
KXY is less than one but greater than zero it is an indication
that output yt is being produced by input xt as well as by
other inputs. Hence, the coherence is nothing else than the
R2 in the frequency domain. Since we are calculating the
coherence using the short time Fourier transform, the
coherence may also be time-varying. So we have to extend
2
KXY by a time index. For the rest of this paper, we will
2
writeKXY ;t .
Suppose now we are interested in the relationship
between two variables fyt g and fxt g, where fyt gis the
wine growth rate and fxt g is the temperature variability for
example. We assume that they are related in the following
way:
 
V ðLÞt yt ¼ AðLÞt xt þ ut ; ut  i:i:d: 0; s 2 ð2:3Þ
where A(L)t and V(L)t are filters, and L is the lag operator
such that Lzt = zt−1. Notice that the lag structure, A(L)t, is
time-varying. That means we need to use a state space
model (we use the Kalman filter again) to estimate the
implied lag structure. That is
 
vi;t ¼ vi;t1 þ "i;t ; for i ¼ 1;:::; p and "i;t  0; s 2"i ð2:4Þ
 
ai;t ¼ ai;t1 þ hi;t ; for i ¼ 0;:::; q and hi;t  0; s 2hi
As before, we test for the random walk property using
the LaMotte–McWorther test,and for structural breaks, we
we use our previous general to specific approach to
estimate (Eq. A.3; Appendix), starting off with lag lengths
of nine and p = q, and dropping those lags which were
never significant (as we did before).
Having estimated the coefficients in Eq. 2.3, we can
calculate the gain, coherence and cross-spectra based on the
time-varying spectra just obtained. This allows us to
overcome a major difficulty in this kind of analysis: namely,
that a very large number of observations would usually be
necessary to carry out the necessary frequency analysis by
362
direct estimation. That would be a particular problem in the
case of structural breaks, since the sub-samples would
typically be too small to allow the associated spectra to be
estimated directly.
The coherence is equivalent to the R2 statistic, and the
gain is equivalent to the regression coefficient, impact or
transmission effect of xt on yt, in the time domain. Thus, the
coherence measures, for each frequency, the degree of fit
between xt and yt: equivalently the R2 between each of the
2
corresponding cycles in xt and yt. Hence AðwÞt and KYX ;t
(see Appendix) measure the link between two variables at
time t. For example, if the coherence has a value of 0.6 at
frequency 1.2, then it means that the temperature cycle at
frequency of 1.2 determines wine production cycle at that
point in time by 60%. Similarly, a gain of 0.5 means that
half the variance in temperature cycle at that frequency is
transmitted to the wine production cycle. In this paper, we
are concerned with the coherence and gain, not with
measuring the phase shift elements as such. But we are
able to detect changes in phase relationships from changes
in the relative importance of different cycles in the cross-
spectral components.
Results
In the figures shown in the next sections, we first present
the time-varying spectrum and then the coherence and
gain.2
Single spectra of Douro wine production
Figure 2 shows the time series of wine production in the
Douro Region from 1940 to 2008. For most of the sample,
this time series shows a lot of variation, which can be
caused by structural breaks. In any case, this variation
makes a common regression very difficult, as it does not
really capture the variation. In contrast, time-varying
parameter approaches can capture those parameter changes
in a systematic way.
Table 1 shows the regression results for the series
production. This AR(5) model is the basis for the spectrum
shown in Figs. 3 and 4. As one can see the regression is
robust as there is no autocorrelation. For the chosen model,
this was in fact, the lowest AIC value. The adjusted Radj-
2 test of the prediction power of our model. We would expect
One can see from the figures that the spectra change. However, one
cannot infer directly from those figures that the changes in the spectra
are also statistically significant. The figures for the time-varying
spectra/cross-spectra have to be accompanied by the fluctuation test
results. Once a structural break has been identified by the fluctuations
test, the results will show up as a significant change in the associated
spectrum or coherence or gain. The results of the fluctuation tests are
available from the authors upon request.
Int J Biometeorol (2012) 56:357–370
squared is relatively high with 75%, but there is a lot of
unexplained variance. Although the first four lags are
statistically not significant at the end of the sample, they
were at other sample points in time, which is why we kept
them in the regression (Table 1). Hence, this table only
shows the final regression for the last observation for
reason of restricted space.
The regressed series models the variation of the
original series fairly well (Fig. 3). But the peaks are
sometimes not as high as in the original series. This calls
for an investigation of other determinants of the dynamic
behaviour of the series. However, before we come to that,
it is worthwhile to highlight the dynamic properties of the
wine production series.
The time-varying spectrum, which is based on above
regression, shows the dynamic characteristics of the wine
production. Over the entire frequency band, there are three
distinctive peaks: at 0.1, 1.3 and 2.5 (Fig. 4). A frequency
of 0.1 basically represents the long run trend. In Fig. 4, the
trend can be seen in the upper left hand corner. Figure 2
contains an estimated trend and this trend is clearly upward
sloping over the entire sample. It has a spectral mass of
about 1.5 and hence is as important as the shorter cycles
(Fig. 4). This is particularly true for the last 5 years of the
sample. That was not always the case, for example during
the 1990s.3
Hence, currently, wine production is characterised by a
long run trend and two shorter cycles of 4.8 years and
2.5 years, respectively. In other words, wine production is
following an (upwards) trend. To show that the trend is
upward sloping, we regressed the raw data on trend.
The Fig. 5 shows the behaviour of wine production
cycles in the Douro region “excluding trend”. Since the two
cycles of wine production have the same spectral density, i.e.
have the same impact, that makes it difficult to distinguish
between the two when considering the time series only. For
example, the upswing of the 5-year cycle could be
dampened by the downturn of the 3-year cycle.
The question now is how to determine the production
peaks. Of course, one could always use some sophisticated
time domain method, but we are looking for something
more practicable. Assume we are interested in the 5-year
cycle.4 In order to create a time series that inhibits the 5-
year cycle, we subtracted from our raw data the trend and
the 3-year cycles. Figure 6 shows the resulting data.
With this series, we can now perform an out-of-sample
a maximum production every 5 years. Counting the
maximum values shows that, of 17 maximum values we
3
The link between period (P) and frequency (ω) is P=2×π/ω.
4
Of course, we could do the same analysis for the trend and the 3-
year cycle.
Int J Biometeorol (2012) 56:357–370 363

Fig. 2 Time series and estimated 2000

linear trend of wine production y = 9.19x + 715.16
R² = 0.41

Wine production (hL * 1000)

1750

1500

1250

1000

750

500

1932
1935
1938
1941
1944
1947
1950
1953
1956
1959
1962
1965
1968
1971
1974
1977
1980
1983
1986
1989
1992
1995
1998
2001
2004
2007
predicted 12 correctly. So for a farmer to make a forecast characterised by a relatively high level of Tm_Sp variation
based on the cyclical properties of wine production there is (mean 16.0°C; σ 0.76); (2) the next period is from 1984 to
a 71% probability that the farmer will make the right 1994, where the variance is relatively low (mean 16.6°C; σ
prediction. 0.47); and (3) in the last regime from 1994 to 2008, the
variation is increasing again but also the Tm_Sp is higher
Cross-spectra of Douro wine production (mean 17.9°C; σ 0.76) than in the previous periods. Given
these structural breaks, it makes sense to incorporate
Having established what characterises the production Tm_Sp into our AR(5) model in a time-varying manner.
cycles, the next question is what causes them. The spectrum Table 2 shows the regression results for the final point in
in itself cannot answer this question. It is purely descriptive. time (2008). In comparison to Table 1, the Radj2 is now
We had the choice of several exogenous variables which much higher (98%) and the AIC value is lower.
may have an impact on wine production. Notably, we have Figure 8 shows the behaviour of the Tm_Sp coefficient.
time series of rainfall and temperature for the period 1967 As one can see, the coefficient varies a lot throughout the
to 2008. The aim was to find a determinant that can explain sample. However, towards the end of the sample, the
the observed production pattern and therefore the “most coefficient stabilises on a relatively high level (149).
important” variable. As it turned out among all models Although, our model does not say what exactly causes the
containing different variables, the one that produced the increase of the impact of temperature on wine production,
lowest AIC value was the one containing mean temperature we argue that due to global warming temperature (IPCC
in spring (Tm_Sp). 2007) becomes a more important factor for wine produc-
Figure 7 show the time series of Tm_Sp and, like the tion. At the end of the sample, a 1°C increase in the Tm_Sp
wine production series, it contains large variation. Visually, will increase mean wine production in the same year by
it seems there are three regimes: (1) one regime that starts 13.4% (149,238 hL; Table 2). Over time though, after
the beginning of the sample until 1983; this period is 3 years, an increase in Tm_Sp will reduce mean wine

Table 1 Kalman filter parameter estimates and summary statistics for time-varying spectral model of the wine production in the Douro Region

Dependent variable PRD Annual data from 1940–2008 Variable Coeff. SE t values

Obs 69 df 64 Constant 363.712 29.88 12.17

Radj2 0.75 SE y 313.75 PRD(1) 0.1123 0.092 1.22
Mean 1,110.87 SSr 5,086,005 PRD(2) 0.0595 0.124 0.48
SE 281.90 LjB test 33.05 PRD(3) 0.0907 0.109 0.83
AIC 325.53 PRD(5) 0.3594 0.150 2.40

Obs Usable observations, df degrees of freedom, PRD Douro wine production, Radj2 adjusted coefficient of determination, AIC Akaike criterion,
LjB test Ljung-Box Test: Q×(16), SSr sum of squared residuals, SE y standard error of dependent variable and of estimate (SE)
t values Significance levels for 64 df: t>1.670 (p<0.05); t>2.387 (p<0.01); t>3.442 (p<0.0005)
364
Fig. 3 Original time series
(black) and regressed time
series (dotted) of wine produc-
Wine Production (hL * 1000)
tion in Douro Region
production by 6.0% (66,279 hL, Table 2).So the increase of
the coefficient has “positive” effects as well as “negative”.
Positive effects arise because higher temperature allows for
higher wine production in the short term. In the long term
though, higher temperatures lead to drier soils which lead to
lower wine production.
The original idea to use Tm_Sp was to check whether this
variable can explain the wine production cycles. Figure 9
shows the gain. What this figure is showing is that if the
temperature changes it causes two major cycles: on the one
hand, there is the cycle at a frequency of 1.2 or 5.24 years,
and on the other, a frequency of 2.6 or 2.41 years. Those
cycles are indeed very close to the production cycles.
Although the two most important cycles of the gain have
always been the most important cycles, their impact is not
constant as one can see from Fig. 9. Nevertheless, the 5.24-
year cycle has always been the most important one. This
result is in contrast to the previous results where the
Fig. 4 Time-varying spectrum
of the wine production
Frequency
Int J Biometeorol (2012) 56:357–370
Year
spectrum of the wine production was characterised by an
equal impact of these cycles. Spring temperature on its own
does not cause cycles of equal strength. Hence, temperature
alone cannot fully explain the dynamic behaviour of wine
production. Other impacts could come from meteorological
variables not tested, grape prices or market regulating
mechanisms, for example.
However, looking at Fig. 9, there is a stable impact on
temperature on wine production. In Fig. 10, the coherence
reveals that for some cycles (frequencies of 3 and 1.2) the
correlation has since 1995 been on a higher level than
before (if one omits the spike in 1986). We can therefore
conclude that correlation between temperature and wine
production is increasing for these cycles. What makes this
result important is that, without having decomposed the
cyclical behaviour, this effect would not have been visible.
If we had looked at only the time domain results, this effect
would have been “averaged out”.
Power
Time
Int J Biometeorol (2012) 56:357–370
Fig. 5 Graphical interpretation of the wine production cycles in
Douro region. There are two fluctuations of equal strengths: one has a
length of 4.8 years (solid line) and there is a shorter one of 2.5 years
(dashed line)
Having established that the spring temperature causes a
dynamic behaviour of wine production close to the spectrum,
the next question is: what can we learn from the actual
behaviour of the temperature? For that, we consider the
coherence which gives the current impact of temperature on
wine production.
The coherence shows that Tm_Sp explains up to 50% of
the 5-year cycle of wine production and about 40% of the
current short run trend of production. Spring temperature
also explains about 20% of the current 2.4-year cycle. The
conclusion is twofold: temperature does have a contempo-
rary effect on wine production. However, this impact is no
more than 50%. Given that the contemporary effect is rather
low, there must be other variables which explain the
contemporary behaviour of wine production.
However, given that wine producers may be interested in
the predictive power of temperature, further research is
needed on how strong this predictive power is.
Discussion
From 1932 to 2008, the growing season for wine production
in the Douro region has been subjected to cycles, but there has
Fig. 6 Time series of wine
Deviation from the trend (hL x 1000)
production filtered for the trend
and the 3-year cycles
365
been a trend toward a continued increase in production. While
some of the trend in high production can undoubtedly be
attributed to vineyard structure changes and/or better viticul-
tural and wine practices, climate exerts an influence on
production variability. As previously stated, the linear model
currently used by industry analysts accounts for less than 50%
of the variation in grape production; our model based on
spring temperature explains approximately 98% of wine
production variability.
According to this time-varying model, the Tm_Sp is the
meteorological variable which explains more of the
cyclicality of Douro wine production. Despite the great
importance of rainfall on grapevine production in Mediter-
ranean climates (Quiroga and Iglesias 2009; European
Commission 1997; Santos et al. 2010), mainly during the
spring and summer period, no significant influence on wine
production cycles was reported in this study. In the Douro
region, summer rainfall is consistently low over the years,
and it is difficult to consider it as an important factor to
explain the great variability of wine production (Cunha et
al. 2003). On the other hand, high spring temperature is
usually combined with high sunshine levels, low precipita-
tion levels and soil moisture.
Remarkably, spring temperature can be used to model
the actual short-term variation of wine production. Mean
spring temperature explains up to 35% of short-term
variation of wine production (Fig. 8). Moreover, it also
explains up to 37% of the current 5-year cycle and it still
explains 15% of the current 2-year cycle (Fig. 10). Hence,
if one follows the Tm_Sp, this information can be used to
establish the stage the wine production cycle. This
information is useful when it comes to forecasts concerning
wine production in the current year.
These wine cycles are consistent with Esteves and
Orgaz (2001) for another wine region of Portugal (Dão).
These authors found, for the period 1960–1992, that
temperature in May explained the 2.5- and 5.3-year wine
cycles in the Dão region, frequencies very similar to the
1500
5 Year Cycle
1000
500
-500
-1000
-1500
366 Int J Biometeorol (2012) 56:357–370

Fig. 7 Mean spring temperature

in Douro region for the period
1973–2008

Temperature ºC
Year

Douro wine cycles explained by the Tm_Sp.Furthermore,
this spectra of wine series in the Dão region showed
statistically significant oscillations coherent with those
found in the series of teleconnection indices.Future
observational research should be aware of the relationships
between climate variability result from the impact of large-
scale geophysical mechanisms and the cyclical behaviour
of wine production.
The influence of previous and in-season spring temper-
ature on the short-term wine cyclical production found in
this study is consistent with previous studies on the
grapevine ecophysiology behaviour and yield. As a
perennial and deciduous plant, environmental conditions
influence its vegetative and reproductive growth. The
conditions of the previous year are important components
of the wine production variability. This influence relates
not only to the season in which the crop is produced but
also to past seasons, mainly the two seasons before the
one in which the harvest takes place (Cunha et al. 2010;
Vasconcelos et al. 2009). As one of the premises for grape
yield, grapevine bud fruitfulness has been the focus of
many studies. Most of these studies have consistently
determined that light and temperature are the most
important climatic factors during the bud differentiation
in the season before the one in which the harvest takes
place (May 2004; Cunha et al. 2010). These studies
support the positive impact of Tm_Sp of the current 2-year
production cycle found in our work.
Hot spring temperatures are favourable, directly or
indirectly, for photosynthetic assimilation and grapevine
pollination/fertilisation, and result in high fruit-set (Cunha
et al. 2003; Vasconcelos et al. 2009). Moreover, hot and
dry spring conditions are generally associated with low
impacts of grape diseases on production level. On the
other hand, the earlier phenology development, associat-
ed to hot springs (Jones and Davis 2000; May 2004),
allows the grapevine to be rain-fed and to capitalise on
diminishing soil water from winter rains. All these effects
are in accordance with positive impact of in-season spring
temperature on the Douro wine production level observed
in our study.
However, in the Mediterranean climate, no irrigated
vineyard (like the Douro region) submitted to consecutive
hot springs could reduce the soil moisture as well as
depleting accumulated reserves of carbon and nutrients in
the permanent structures (e.g. May 2004). These reserves
can play a critical role in potential production of grapevine,
mainly in years with unfavourable climate conditions
during the spring (García-de-Cortázar-Atauri 2006). Mod-
elling work suggests a negative relationship between wine
production and the spring temperatures in the previous
5 years.

Table 2 Kalman filter parameter estimates and summary statistics for the time-varying model of wine production on temperature

Dependent variable PRD Annual data from 1973–2008 Variable Coeff. SE t values

Obs 36 df 32 Constant −502.368 19.645 −25.57

R2ad 0.98 SE y 313.75 PRD(5) 0.3015 0.0275 10.96
Mean 1,110.87 SSr 2,040,501 Tm_Sp 149.238 9.7644 15.28
SE 252.52 LjB test 12.159 Tm_Sp(3) −66.279 8.5834 −7.72
AIC 315.35

Obs Usable observations, df degrees of freedom, PRD Douro wine production, Radj2 adjusted coefficient of determination, AIC Akaike criterion,
LjB test Ljung-Box Test: Q×(16), SSr sum of squared residuals, SE y standard error of dependent variable and of estimate (SE)
t values Significance levels for 64 df: t > 1.695 (p<0.05); t>2.452 (p<0.01); t >3.643 (p<0.0005)
Int J Biometeorol (2012) 56:357–370
Fig. 8 Immediate impact of
spring temperature
Parameter Value [ ]
Modelling cyclical wine production with this time-varying
model allows for accurate short-term as well long-term
forecasts, providing long-term growth rates and short-term
cycle lengths and intensities by different periods and climate
scenarios.The projections of future climate in combination
with the developed model of cyclical behaviour of wine
production can provide a powerful tool for assessing potential
future responses of viticulture and the wine industry to
climatic variability, and allow statistical confidence limits to
be attached to estimated responses. With respect to the future,
we can expect climate risks to intensify in the Mediterranean
(IPCC 2007) which represent an additional challenge for
wine production in the Douro region. Based on the most
recent and comprehensive ensembles of global and regional
climate variability simulations, the Mediterranean may
experience substantial warming (temperature increases of
3–5°C) by 2080; at the same time, inter-annual variability is
projected to increase, especially in the spring period (Giorgi
and Lionello 2008; IPCC 2007).According to our model,
Fig. 9 Gain: temperature on
production
367
Year
wine production would increase by 22% if temperatures rise
by 3°C. This result is in line with Santos et al. (2010) who
predict an increase in wine production of approximately
25% for the Douro region.
Viticultural regions, which produce premium wines, are
usually located in narrow climatic niches (Bindi et al. 1996,
2001; Jones 2007). These climatic niches are very vulner-
able to both short-term climate variability and long-term
climate change. This in turn may cause adverse effects on
the viticultural regions.
Due to the economic and social importance of the wine
industry (EU 2006; OIV 2010), it behoves us to pay
attention to this phenomenon, and especially to put effort
into understanding how viticulture and the wine industry
will respond and adapt to information about climate
generally and modelling in particular. Our model suggests
that stationarity is a questionable assumption, and this
means that historical distributions of wine production are
going to need dynamic updating. We should consider this
Power
Time
Frequency
368
Fig. 10 Coherence between
wine production and spring
temperature
when developing wine production models where climate
plays an important direct or indirect role and climate
variability is increasing, as indicated by the recent IPCC
reports.
Acknowledgements The authors would like to thank the Instituto
dos Vinhos do Porto e Douro (IVDP) for providing the wine
production data and the Associaçãopara o Desenvolvimento da
ViticulturaDuriense (ADVID-Douro Region Wine Cluster) for pro-
viding the climate data and two anonymous referees for their very
useful comments.
Appendix: The Short Time Fourier Transform (STFT)
In discrete time, this means data to be transformed has
been broken up into frames (which usually overlap each
other). Each frame is then Fourier transformed, and the
(complex) result added to a matrix which records its
magnitude, phase and frequency at each point in time.
This can be expressed as:
X
1
STFT fx½ng  X ðm; wÞ ¼ x½nw½n  mejwn ðA:1Þ
n¼1
In this case, m and n are different points in time; ω is
the frequency and is continuous; and j=√-1. But in most
typical applications ,the STFT is performed using the Fast
Fourier Transform, so all variables are discrete and “n-m”
would be the estimation window. In our application, the
window is not constant, but increasing with each new
observation. Moreover, we derive the STFT using Kalman
filter estimates of Eq. 2.1; see “Cross-spectrum analysis”
Int J Biometeorol (2012) 56:357–370
above. The squared magnitude of the STFT then yields the
spectrogram of the function:
Spectrogram f# t g  jX ðt; wÞj2 ðA:2Þ
The remaining question is what algorithm do we use to
calculate the Fast Fourier Transform? One algorithm often
used to calculate the Fast Fourier Transform is the Bluestein
algorithm (Bluestein 1968), which is also called the chirp z-
transform algorithm .In particular, it can compute any
transform of the form:
X
N 1
Xk ¼ xn znk ; where k ¼ 0; :::; M  1 ðA:3Þ
n¼0
for an arbitrary complex number z and for differing
numbers N and M of inputs and outputs (see also Rabiner
et al. 1969). Hence, the algorithm we apply to calculate the
Fast Fourier Transform is a well-established algorithm and
widely used in engineering (Boashash and Reilly 1992;
Boashash 2003). It is not commonly used in economics,
however.
Finally, Boashash and Reilly (1992) have shown
theoretically that, once Eq. 2.2 (see “Cross-spectrum
analysis” above) has been estimated, its coefficients αi,t
can be used to calculate the short time Fourier Transform
and the power spectra directly (by applying the Bluestein
algorithm). That has the convenient property that the
traditional formulae for the PSD are still valid and may
still be used, but they have to be recalculated at each
point in time. The time-varying spectrum of the growth
rate series can therefore be calculated as follows (see
Int J Biometeorol (2012) 56:357–370 369

also Lin 1997): References

s2
Pt ðwÞ ¼  2 ðA:4Þ Artis M, Marcellino M, Priorietti T (2004) Dating the euro area
 
1 þ P ai;t exp ðjwiÞ
9
business cycle. In: Reichlin L (ed) The Euro area business cycle:
  stylised facts and measurement issues. Centre for Economic
i¼1 t
Policy Research, London
Bindi M, Fibbi L, Gozzini B, Orlandini S, Miglietta F (1996)
where ω is angular frequency and j is a complex number. Modelling the impact of future climate scenarios on yield and
The main advantage of this method is that, at any point in yield variability of grapevine. Climate Res 07(3):213–224
time, a power spectrum can be calculated instantaneously Bindi M, Fibbi L, Miglietta F (2001) Free Air CO2 Enrichment
(FACE) of grapevine (Vitis vinifera L.): II. Growth and quality of
from the updated parameters of the model. Hence, we are grape and wine in response to elevated CO2 concentrations. Eur J
able to generate a power spectrum even if we have a short Agron 14(2):145–155
time series and even if that time series contains structural Bluestein LI (1968) A linear filtering approach to the computation of
breaks. the discrete Fourier transform. Northeast Electronics Research
and Engineering Meeting Record 10:218–219
For the crossspectral analysis, we use the methods Boashash B (ed) (2003) Time frequency signal analysis and process-
introduced in Hughes Hallett and Richter (2009a, b, c). ing. Elsevier, Oxford
The time-varying cross spectrum, fYX(ω)t, using the STFT Boashash B, Reilly A (1992) Algorithms for time-frequency signal
can be written as: analysis. In: Boashash B (ed) Time-frequency signal analysis -
methods and applications. Longman-Cheshire, Melbourne, pp
fYX ðwÞt ¼ jT ðwÞjt fXX ðwÞt ðA:5Þ 163–181
Canova F (1998) Detrending and business cycle facts. J Monetary
where T(ω)t is the transfer or filter function is defined by Econ 41:475–512
Canova F, Dellas H (1993) Trade dependence and the international
Eq. A.5 and calculated as follows:
business cycle. J Int Econ 34:23–47
Capitanio F, Adinolfi F (2009) The relationship between agricultural
0 P q 1 insurance and environmental externalities from agricultural input
ab;t exp ðjwbÞ
B b¼0 C use: a literature review and methodological approach. New Medit
T ðwÞt ¼ B
@
C; for t ¼ 1;:::; T
A ðA:6Þ 8(3):41–48
P
p
Casa-do-Douro (1999) Dados estruturais da região do Douro.
1  vi;t exp ðjwiÞ
i¼1 Cadastro da casa do Douro. Douro-Boletim da Casa do Douro,
Peso-da-Régua, Portugal
Chaves M, Rodrigues L (1987) Photosynthesis and water relations in
The last term in Eq. A.5, fXX(ω)t, is the spectrum of grapevines response to environmental factors. In: Tenhunen JD
predetermined variable. This spectrum may also be time- (ed) Plant Response to Stress-functional analises in Mediterra-
varying. However, in this paper, we are interested in the nean ecosystems. Springe, Berlin, pp 279–290
Chen C, Chang C (2005) The impact of weather on crop yield
coherence and in the composition of the changes to that distribution in Taiwan: some new evidence from panel data
coherence over time. So we need to establish expressions models and implications for crop insurance. Agric Econ 33:503–
for the coherence and gain between xt and yt to show the 511
degree of association and size of impact of xt on yt. The Chen CC, McCarl BA, Schimmelpfennig DE (2004) Yield variability
as influenced by climate: a statistical investigation. Clim Change
spectrum of any dependent variable is defined as (Wolters 66(1–2):239–261
1980; Nerlove et al. 1995; Jenkins and Watts 1968; Laven Corte-Real J, Qian BD, Xu H (1998) Regional climate change in
and Shi 1993): Portugal: precipitation variability associated with large-scale
atmospheric circulation. Int J Climatol 18(6):619–635
 2
fYY ðwÞt ¼ T ðwÞt  fXX ðwÞt þ fvv ðwÞt ðA:7Þ Cunha M, Abreu I, Pinto P, De Castro R (2003) Airborne pollen
samples for early-season estimates of wine production in a
Mediterranean climate area of northern Portugal. Am J Enol Vitic
From Eq. A.4, we get the time-varying residual spectrum
54(3):189–194
fuu ðwÞt Cunha M, Marçal ARS, Silva L (2010) Very early prediction of wine
fvv ðwÞt ¼  2 ðA:8Þ yield based on satellite data from VEGETATION. Int J Remote
 
1  P vi;t exp ðjwiÞ
p
Sens 31(12):3125–3142
  Esteves MA, Orgaz MDM (2001) The influence of climatic variability
i¼1
 2 on the quality of wine. Int J Biometeorol 45(1):13–21
and the gain as AðwÞt ¼ T ðwÞt  : Finally, given knowledge EU (2006) Wine, Economy of the sector. European Commission,
of fYY(ω)t,T ðwÞt  , and fXX(ω)t, we can calculate the
2 Directorate-General for Agriculture and Rural Development.
Available online at http://ec.europa.eu/agriculture/markets/wine/
coherence at each frequency as: studies/rep_econ2006_en.pdf (assessed 11 November 2009)
European Commission (1997) Oliwin Project: agrometeorogical
models for the estimation at harvest of olive and vine yield;
1
2
KYX ;t ¼
n . 2 o ðA:9Þ regional and national level. Final Report
1 þ fVV ðwÞt T ðwÞt  fXX ðwÞt Ferris J (2006) Forecasting World Crop Yields as a Probability
Distributions. Paper presented at the International Association of
370
Agricultural Economists Conference, Gold Coast, Australia, 12–
18 August
Folwell RJ, Santos DE, Spayd SE, Porter LH, Wells DS (1994)
Statistical technique for forecasting concord grape production.
Am J Enol Vitic 45(1):63–70
Fonseca A (1949) O beneficio e a sua distribuição na Região
Vinhateira do Douro. Federação dos Viticultores da Região do
Douro, Peso-da-Régua, Portugal
Gabor D (1946) Theory of communication. J Inst Electr Eng 93
(3):429–457
García-de-Cortázar-Atauri I (2006) Adaptation du modèle STICS à
la vigne (Vitis vinifera L.). Utilisation dans le cadre d’une
étude du changement climatique à l’échelle de la France. PhD
thesis, Ecole Supérieur Nationale d’Agronomie de Montpellier,
Montpellier
Gemmil G (1978) Estimating and forecasting agricultural supply from
times-series: a comparison of direct and indirect methods. Eur
Rev Agric Econ 5(2):175–191
Giorgi F, Lionello P (2008) Climate change projections for the
Mediterranean region. Glob Planet Change 63(2–3):90–104.
doi:10.1016/j.gloplacha.2007.09.005
Goodwin BK, Ker AP (1998) Nonparametric estimation of crop yield
distributions: implications for rating group-risk crop insurance
contracts. Am J Agric Econ 80(1):139–153
Hughes Hallett A, Richter C (2009a) Economics in the backyard: how
much convergence is there between China and her Special
Regions? World Econ 32(6):819–861
Hughes Hallett A, Richter C (2009b) Has there been any structural
convergnence in the transmission of european monetary policies?
Int Econ Econ Policy 6(2):85–101
Hughes Hallett A, Richter C (2009c) Is the US No longer the
economy of first resort? Changing economic relationships in the
Asia-Pacific Region. Int Econ Econ Policy 6(2):207–234
Hurrell JW (1995) Decadal trends in the North Atlantic Oscillation:
regional temperatures and precipitation. Science 269(5224):676–
679. doi:10.1126/science.269.5224.676
Hurrell JW, Van Loon H (1997) Decadal variations in climate
associated with the North Atlantic Oscillation. Clim Change 36
(3):301–326. doi:10.1023/a:1005314315270
IPCC (2007) Climate Change (2007). Fourth Assessment Report of
the Intergovernmental Panel on Climate Change. Cambridge
University Press
IVDP (2010) Instituto dos Vinhos do Douro e Porto, dados estatísticos
sobre a produção de vinho do Douro e Porto na Região
Demarcada do Douro
Jenkins GM, Watts DG (1968) Spectral analysis and its applications.
Holden-Day, San Francisco
Jones GV (2007) Climate Changes and the global wine industry. Paper
presented at the 13th Australian wine industry technical
Conference, Adelaide, Australia, 28 July – 2 August
Jones GV, Davis RE (2000) Climate influences on grapevine
phenology, grape composition, and wine production and quality
for Bordeaux, France. Am J Enol Vitic 51(3):249–261
Knippertz P, Ulbrich U, Marques F, Corte-Real J (2003) Decadal
changes in the link between El Nino and springtime North
Int J Biometeorol (2012) 56:357–370
Atlantic oscillation and European-North African rainfall. Int J
Climatol 23(11):1293–1311. doi:10.1002/joc.944
LaMotte LR, McWorther AJ (1978) An exact test for the presence of
random walk coefficients in a linear regression. J Am Stat Assoc
73(364):816–820
Laven G, Shi G (1993) Zur Interpretation von Lagverteilungen.
Discussion Paper, Johannes Gutenberg University, Mainz
Lin Z (1997) An Introduction to time-frequency signal analysis.
Sensor Rev 17(1):46–53
May P (2004) Flowering and Fruitset in Grapevines. Lythrum Press,
Australia
McCarl BA, Villavicencio X, Wu XM (2008) Climate change and
future analysis: is stationarity dying ? Am J Agric Econ 90
(5):1241–1247. doi:10.1111/j.1467-8276.2008.01211.x
Milly PCD, Betancourt J, Falkenmark M, Hirsch RM, Kundzewicz
ZW, Lettenmaier DP, Stouffer RJ (2008) Climate change -
stationarity is dead: whither water management? Science 319
(5863):573–574. doi:10.1126/science.1151915
Nerlove M, Grether DM, Carvalho JL (1995) Analysis of economic
time series. Academic, New York
OIV (2010) Organization Internationale de la Vigne et du Vin -
Statistiques. Accessed 1 January 2010
Paredes D, Trigo RM, Garcia-Herrera R, Trigo IF (2006) Understand-
ing precipitation changes in Iberia in early spring: weather typing
and storm-tracking approaches. J Hydrometeorol 7(1):101–113.
doi:10.1175/JHM472.1
Phares C (2000) Business cycles and the world wine market. Regional
Research Institute, West Virginia University
Ploberger W, Krämer W, Kontrus K (1989) A new test for structural
stability in the linear regression model. J Econ 40(2):307–318
Quiroga S, Iglesias A (2009) A comparison of the climate risks of
cereal, citrus, grapevine and olive production in Spain. Agric
Syst 101(1–2):91–100. doi:10.1016/j.agsy.2009.03.006
Rabiner LR, Schafer RW, Rader CM (1969) The chirp z-transform
algorithm and its application. Bell Syst Tech J 48:1249–1292
Reis R, Lamelas H (1988) Statistical study of decade series of water
balance and its components of potencial evapotranspiration
calculated by Penman's method. vol 36. Instituto Nacional de
Meteorologia e Geofisica, Lisbon
Santos JoA, Malheiro AC, Karremann MK, Pinto JG (2010) Statistical
modelling of grapevine yield in the Port Wine region under
present and future climate conditions
Vasconcelos MC, Greven M, Winefield CS, Trought MCT, Raw V
(2009) The flowering process of vitis vinifera: a review. Am J
Enol Vitic 60(4):411–434
Vicente-Serrano SM, Heredia-Laclaustra A (2004) NAO influence on
NDVI trends in the Iberian peninsula (1982–2000). Int J Remote
Sens 25(14):2871–2879. doi:10.1080/01431160410001685009
Wells C (1996) The kalman filter in finance vol. 32. Advanced studies
in theoretical and applied econometrics. Kluwer, Dordrecht
Wolters J (1980) Stochastic dynamic properties of linear econometric
models. Springer, Berlin
Zhu Y, Goodwin B, Ghosh S (2008) Time-varying yield distribuition and
the implications for crop insurance pricing. Paper presented at the
NC State Agricultural Economics Workshop, Raleigh, NC, 6 May